MONTE CARLO METHODS

IN A NUTSHELL
A short course
MALCOLM SAMBRIDGE
RESEARCH SCHOOL OF EARTH SCIENCES
OCTOBER 2012

SHORT COURSE STRUCTURE

LECTURES
11.00AM - 12.30PM LECTURE 1, JAEGER SEMINAR ROOM

2.00PM - 3.15PM LECTURE 2 “ “ “

11.00AM - 12.30PM LECTURE 3, “ “ “

2.00PM - 3.15PM LECTURE 4 “ “ “

PRACTICAL AND DISCUSSION TIME

ANU 2.00PM - 5.00PM MONDAY 15TH OCTOBER J8 SEMINAR ROOM

GA 2.00PM - 5.00PM TBA

LECTURE NOTES AND OTHER RESOURCES

http://rses.anu.edu.au/~malcolm/Teaching/MC/

COURSE OVERVIEW

A SHORT HISTORY OF MONTE CARLO METHODS

SOLVING PROBLEMS WITH RANDOM NUMBERS

UNCERTAINTY ESTIMATION AND THE BOOTSTRAP

FITTING PARAMETERS TO DATA WITH STOCHASTIC

METHODS

RESEARCH DIRECTIONS: MCMC SAMPLING

3
 AUSCOPE INVERSION LAB
The AuScope ilab is a component of the Australian Geophysical Observing System (AGOS)
for development and distribution of advanced data inference software tools to the geoscience
community (2011-2014).

Inference software currently available/under development through ilab:

• Parallelized uniform nested grid search in hyper cube and hyper sphere
•Direct search in multi-dimensional parameter space
(Neighbourhood algorithm, Parallel Tempering)

•General Bayesian sampling tools

•Trans dimensional regression and general inference for 1-D and 2-D spatial domains
•Access to non locally produced inference software packages.

- ilab
ilab Software

SOME REFERENCES
Integrals and data fitting
NUMERICAL RECIPES (3RD ED.) CHAPTERS 7 & 10.
PRESS, W. H., TEUKOLSKY, S. A., VETTERLING, W. T. AND FLANNERY, B. P.,
CAMBRIDGE, 2007. HTTP://WWW.NR.COM

Error analysis
BOOTSTRAP METHODS AND THEIR APPLICATIONS.
DAVISON, A. C.; HINKLEY, D. (2006).
CAMBRIDGE: CAMBRIDGE SERIES IN STATISTICAL AND PROBABILISTIC MATHEMATICS. SOFTWARE.

APPLIED RESAMPLING METHODS - COURSE NOTES (2001, ANU)

MICHAEL MARTIN, STEVEN STERN (AVAILABLE AS PDF FROM SPEAKER)

Data fitting
MONTE CARLO ANALYSIS OF INVERSE PROBLEMS
MOSEGAARD, K., AND SAMBRIDGE, M., INVERSE PROBLEMS, 18, R29-R54, 2002.

MONTE CARLO METHODS IN GEOPHYSICAL INVERSE PROBLEMS

SAMBRIDGE, M., AND MOSEGAARD, K., REV. GEOPHYS., 40, DOI:10.1029/2000RG000089 5 DECEMBER 2002.

There are many resources around. Google and wikipedia can be a good place to start.

HEALTH WARNING:
THIS SHORT COURSE MAY CONTAIN
SOME MATHEMATICS
'There was a tale about a mathematician who went slightly mad, and was greatly affected by the moon.
He was so madly obsessed with the moon, that he made a beautiful model of it, and then he became
convinced that his model was the real moon, and the thing in the sky merely a figment of the
imagination ! Need I say more than "Mathematicians beware of being seduced by your beautiful models".

A. T. Price, The Theory of Geomagnetic Induction, PEPI, vol 7, 227-233, 1973.

‘Beware of mathematicians, and all those who make empty prophecies. The danger already exists that
the mathematicians have made a covenant with the devil to darken the spirit and to confine man in the
bonds of Hell.’
St. Augustine, (AD 354 - 451) De Genesi ad Litteram

‘I think that most of us have met the type of mathematician who is so dazzled by the beauty of his
mathematics that he applies it blindly to all and sundry without adequate analysis of the premises on
which he bases his deductions, the type whose mind has been made over-rigid by pure mathematics! An
important feature of applied mathematics is that it tends to correct this type of mind and, when well
taught, to focus needed attention on the problem of initial premises.’

Keith Edward Bullen (1906-1976)

6
 WHAT ARE MONTE CARLO
METHODS ?
A BROAD CLASS OF COMPUTATIONAL METHODS FOR
SOLVING SEVERAL TYPES OF PROBLEM

DETERMINISTIC PROBLEMS LIKE INTEGRALS

SIMULATION OF RANDOM PROCESSES,

E.G. PROPAGATING UNCERTAINTY, ERROR ANALYSIS

FITTING PARAMETERS TO DATA (OPTIMIZATION)

PROPERTIES
CAN BE CONCEPTUALLY SIMPLE BUT COMPUTATIONALLY INTENSIVE.

IN SOME CASES THEY ARE THE ONLY WAY FORWARD (WHEN MATHEMATICAL
SOLUTIONS DO NOT EXIST).

MONTE CARLO
A SHORT HISTORY

WHERE IS THAT?
The Monte Carlo Casino

“A branch of experimental mathematics that is

concerned with experiments on random numbers.”
Hammersley & Handscomb (1964)

9
 MONTE CARLO: HISTORY

‘Monte Carlo’ coined by

Metropolis and Ulam (1949)

Originally developed by Ulam,

von Neumann and Fermi to simulate
neutron diffusion in fissile materials
in the development of the atom bomb

10

MONTE CARLO: HISTORY

But did someone think of it earlier?

“Astonishingly modern Monte Carlo

techniques” were described by Lord
Kelvin in a discussion of how to solve
the Boltzmann equation (1901).
Hammersley and Handscomb (1964)

Describes the statistical distribution of one particle in a fluid

William Thomson, Lord Kelvin (1824-1907)

S = k ln W
http://en.wikipedia.org/wiki/Boltzmann_equation

11

MONTE CARLO: HISTORY

But did someone think of it even earlier?

Hall (1874) recounts numerical

experiments to determine the value
of pi by injured officers during the
American Civil War (1861-1865).

Buffon’s needle problem (1733)

Determine the value of pi by throwing needles !

12
 MONTE CARLO: HISTORY

Monte Carlo received increased focus in the 1950s

when commercial applications were discovered.
Practical successes in operations research.

Thomson (1957) describes a Monte Carlo simulation of

fluctuations of traffic in the British Telephone system.

It is still used to simulate

traffic flow !

...and other traffic problems!

13

APPLICATIONS OF MONTE CARLO

A class of computational algorithm that rely on repeated random
sampling to achieve their results.

SIMULATION OF PHYSICAL AND

MATHEMATICAL SYSTEMS
Fluids, Disordered materials, operations research

CALCULATION OF RISK IN BUSINESS

INTRACTABLE CALCULATIONS SIMULATION OF A RAYLEIGH-TAYLOR

Integrals in Physics, Chemistry INSTABILITY PROBLEM

HTTP://EN.WIKIPEDIA.ORG/WIKI/FILE:INSTABILITY.JPG

PHENOMENA WITH UNCERTAINTY IN INPUTS

-> UNCERTAINTY ANALYSIS

14

MONTE CARLO
SOLVING PROBLEMS WITH
RANDOM NUMBERS

15
 WHAT ARE RANDOM NUMBERS?
Computers generate a sequence of numbers between 0 and 1 which have
statistical properties that approximate a uniform random sequence.

A seed number is required to start the sequence off (usually a large integer)

a = ran(iseed)
Here’s a sequence from a random number generator with seed = 61254557

0.717, 0.03, 0.199, 0.422, ...

Here’s the histogram of numbers generated

500 5000 50000

70
700 7000

60
600 6000
Frequency

Frequency

Frequency
50
500 5000

40
400 4000

30
300 3000

20
200 2000

10
100 1000

x x x
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

16

UNIFORM RANDOM NUMBERS IN

2-D
Random points in a plane (random numbers in two dimensions)

Pseudo - random Quasi - random

Asymptotically space filling, statistically independent from each other

17

BUFFON’S NEEDLE PROBLEM

Posed by Buffon in 1733 (solved by Buffon in 1777)

Given a needle of length l dropped on a plane ruled

with parallel lines a distance t apart, what is the
probability that the needle will cross a line?

Experiment
h red needles cross the line out of n throws

500 throws of the needle are shown

l = t/3 h = 107 n = 500

2ln
= = 3.116 ± 0.073
th
Solution < 1% error
h 2l
= l t Using random numbers we can estimate
n t
⇤ ⇧ ⇥⌅
h 2 t
= l l2 t2 t sin 1
+1 l>t HTTP://WWW.VENTRELLA.COM/BUFFON/INDEX.HTML
n t l
HTTP://WWW.METABLAKE.COM/PI.SWF

18
 SOLVING PROBLEMS WITH
RANDOM NUMBERS
A major use of MC techniques is in solving integrals

What is the probability that my random dart will hit the dart board?

As = Area of square
Ac = Area of circle

h = Number of hits N = Number of throws

Ac h
P = =
As N

19

MONTE CARLO INTEGRATION

A major use of MC techniques is in solving integrals

If
1 x inside circle
f (x) =
I= f (x)dx x R N
0 otherwise

Then I = area of the circle, Ac

Example: throw random darts at a dart board
Monte Carlo approximation
N
As = Area of square f (xi )
IM C
Ac = Area of circle i=1
(xi )
where (x) is the sampling density
Uniform random samples gives
Count the darts that fall inside the circle
N
(xi ) =
The area of the circle can be approximated by MC
As
N
h h = Number of hits As As h
Area of circle ⇥ As IM C f (xi ) =
N N = Number of throws N i=1
N

20

MONTE CARLO INTEGRATION

THE FULL STORY
In general any integral of the form

I= f (x)dx (x)dx = N
D d

a Monte Carlo approximation is

N
fi
IM C + error fi = f (xi ), i = (xi )
i=1 i

where (x) is the known sampling density and the error term is given by

⇤
N N
⇥2 ⌅1/2 Hey presto
⌥ fi2 1 ⌥
⇥M C =⇧
fi ⌃ you can use this
2 N
i=1 i i=1 i
to solve any integral !

As the number of samples, N increases the error in IM C decreases.

21
 MONTE CARLO INTEGRATION
PRACTICAL
The mass of irregularly shaped objected

(1,1)

I= f (x, y)dxdy
R
R1

x = -0.6 (0,0) x x = 0.6

R2

R3

(-1,-1)

With no analytical solution we can get a numerical approximation through Monte Carlo integration

22

WE CAN SAMPLE OTHER

DISTRIBUTIONS TOO!
There are standard methods for generating random numbers whose density follows well known curves,
like a Bell curve (Normal, Gaussian distribution)

1 (x xo ) 2

p(x) = e 2 2

a = gasdev(iseed) ⇥ 2

x
There are many computer codes available to generate Normally distributed random deviates.
Often these computer routines are just generating uniform deviates and transforming them to Normal
deviates. E.g. using the Box-Muller method
Histogram of deviates 7000

x2 6000

x2 = 2 ln y1 sin 2 y2
5000

4000

3000

x1 = 2 ln y1 cos 2 y2 2000

1000

0
-3 -2 -1 0 1 2 3
x1
x
23

MONTE CARLO INTEGRATION

ANOTHER EXAMPLE
Suppose we want to find the area under the curve below between A and B.
This is the integral, I.
B
I= f (x)dx If we generate our samples according
A to a Gaussian sampling density then we
have
N x2
i
⇥(x) = e 2
2
f
N
(x) IM C
fi
i
fi i=1
I
7000

6000

5000

4000

A B
3000

x xi 2000

1000

0
-3 -2 -1 0 1 2 3

There is no mathematical solution but we can get

Samples generated
a numerical approximation through Monte Carlo integration

24
 WHY ALWAYS GAUSSIAN ?
Given five arbitrary but independent probability distributions, p(xi ), (i = 1, . . . , 5)
what is the PDF of their sum ? p(X); X = x1 + x2 + x3 + x4 + x5

100000
X
deviates
The Central
Limit Theorem

0 0.5 1.0 0 0.5 1.0

0 2.5 5.0 X1 X2

0 0.5 1.0 0 0.5 1.0 0 0.5 1.0

25

MONTE CARLO INTEGRATION

B
SUMMARY I= f (x)dx
A

NUMERICAL INTEGRATION WITH THE MONTE CARLO METHOD IS VERSATILE.

WE GET A NUMERICAL ESTIMATE THAT GETS MORE ACCURATE AS MORE

SAMPLES ARE GENERATED.

ERROR ESTIMATES CAN BE OBTAINED AND USED TO DECIDE WHEN TO STOP

SAMPLING.

THE PROCESS IS EFFICIENT IF THE SAMPLING DENSITY IS SIMILAR TO THE

ACTUAL FUNCTION BEING INTEGRATED.

MANY APPLICATIONS ACROSS THE SCIENCES. BUT THE PROCESS BECOMES

COMPUTATIONALLY EXPENSIVE IN HIGH DIMENSIONS. MUCH RESEARCH ON
METHODS FOR CHOOSING (ADAPTING) THE SAMPLING DENSITY (x).

26

MONTE CARLO
UNCERTAINTY PROPAGATION

`He believed in the primacy of doubt, not as a blemish on our ability to know but
as the essence of knowing.'
James Gleick's quote on Feynmann

27
 MONTE CARLO
UNCERTAINTY PROPAGATION
In the Earth Sciences we often do the following

Some data Important

Data
processing result !

For example
The outputs y1 , y 2 may represent the age of a zircon,
uplift rate of a coral reef, location of an earthquake
or anything else.

Given some understanding (or assumptions) of what

Given the errors in x, what are the errors in y ? type of errors are present in the data (x1, ...) we want
to know how they propagate into the estimated parameters.

If f(x) is simple then calculus can solve this problem. But as it gets more complicated Monte Carlo
methods become a powerful alternative.

28

MONTE CARLO
UNCERTAINTY PROPAGATION
The general procedure

STEP 1: CREATE A PARAMETRIC MODEL, y = f (x1 , x2 , . . . , xn ).

STEP 2: GENERATE A SET OF INPUTS WITH RANDOM NOISE, x1 , x2 , . . . , xn .

xi xi + i , (i = 1, . . . , n)
This is where the random deviates are added in

STEP 3: EVALUATE THE MODEL AND STORE THE RESULTS AS yj (j = 1, . . . , m)

STEP 4: REPEAT STEPS 2 AND 3 MANY TIMES

STEP 5: ANALYZE THE RESULTS USING HISTOGRAMS,

SUMMARY STATISTICS, CONFIDENCE INTERVALS, ETC.

29

MONTE CARLO
UNCERTAINTY PROPAGATION
Example estimating the error in an earthquake hypocentre

With travel time data to each seismic station we can calculate a best
fit estimate of the earthquake’s position (x,y,z) and origin time, t.

Travel Least squares Hypocentre

times location procedure (x,y,z,t)

Monte Carlo can be used to determine the uncertainty in the hypocentres

by assuming a statistical distribution for errors in the travel times.
Random errors in travel times are generated according to their respective distributions and added to the data.

t1 t1 + 1
Relocate Repeat many times
t2 t2 + 2
earthquake

... etc.

30
 MONTE CARLO
UNCERTAINTY PROPAGATION
Earthquake hypocentre location: Results of Monte Carlo error simulation

From the distribution of

results we can calculate the
standard deviation or the
hypocentres or
confidence intervals

From Billings et al. (1995)

31

BOOTSTRAP METHODS FOR

UNCERTAINTY ANALYSIS
What if we don’t know the statistics of the errors in the data ?

Data Important
Data
processing result !

Don’t know the

distribution of the
data errors

Errors in Errors in
inputs outputs

In some cases we can use the Bootstrap method

32

THE BOOTSTRAP LEGEND

The bootstrap was invented by Efron (1979) and gets its name because it uses
the variability in the data to estimate the data errors. Its a bit like pulling
oneself up by ones own bootstraps.

Bootstrap
The  travels  and  
surprising  adventures  of  
Baron  Munchausen
Raspe

This idea is often attributed to Raspe’s book about Baron Munchausen (1785),
‘to pull one self up by ones bootstraps’, but he actually did it with his pigtail!

33
 INDEPENDENT IDENTICALLY
DISTRIBUTED (I.I.D.)
Caution: the bootstrap method can ONLY BE APPLIED when the data are
Independent and Identically Distributed random deviates. This means
they are independently drawn from the same probability distribution.

All things being equal the following are examples of IID data
a sequence of outcomes of spins of a roulette wheel is IID.
If a roulette wheel lands on red 20 times in a row it is no more likely to land on
red on the next throw as black. Because the the spins are independent.

a sequence of dice rolls is IID.

a sequence of coin flips is IID.

... or perhaps the age distribution of zircons from a single source

... or your data

Many data are IID or can be transformed to IID

yobs = ytrue +

34

THE BOOTSTRAP METHOD

We have one data set with which to perform our calculation. Wouldn’t it be
just grand if we could repeat the same experiment many times on new data !
Input Output

y = (y1 , y2 , y3 , . . . , yn ) Data Procedure Result X

7000

6000

5000

Data
y1 set 1
Procedure Result 1 X1
4000

3000

2000

1000

0
-3 -2 -1 0 1 2 3

x
X
y2 Data
set 2
Procedure Result 2 X2 Could calculate
mean, ,
...

...

confidence intervals
...

...

...

yM
XM

With many repeat data sets we could look at the distribution of X values and estimate uncertainty !

35

THE BOOTSTRAP METHOD

The bootstrap method is a way of generating new data from your existing data.

Bootstrap principle:
The relationship between uncollected data and your actual data
is the same as between your data and resamples of your data

uncollected collected
Real world data data We don’t have this !

resamples of
collected
Bootstrap world data
collected
data But we do have this !

By using the data itself we can generate ‘new’ data sets by randomly resampling it with replacement

36
 BOOTSTRAP: AN EXAMPLE
Suppose I have five measured ages of my rock, y = (1.1, 2.4, 1.7, 3.1, 4.5)
and the spread in the values is due to some random uncertainty associated with unknown factors
in the measurement process. I decide to take the mean of my data as my best estimate of the age,
5
1
X0 = yi = 2.56. Data Procedure Result
5 i=1

I would like to know how the errors in the data map into this mean. I do not know anything about the
probability distribution of the errors in the data except that each measurement is independent. We
have I.I.D. data and can use the bootstrap!.

We randomly generate copies of the data with replacement

collected resamples of
data data
y1 = (3.1,1.1,1.7,1.1, 2.4) X1 = 1.88 X0

y2 = (2.4, 1.7, 3.1, 3.1, 4.5) X2 = 2.96 Output

y3 = (1.1, 4.5, 1.1, 3.1, 1.7) X2 = 2.3 95%

Confidence
...

...

...

Interval

X
37

BOOTSTRAP FORMULAE
What can you do with the bootstrap samples?

Data Function of the Result

y data f (y) x

With input data y0 = (y1 , y2 , . . . , yn ) and output values xo = (x1 , x2 , . . . , xm )

and N bootstrap data sets generated randomly with replacement yi , (i = 1, . . . , N )

we apply xi = f (yi ) and get B bootstrap samples xi , (i = 1, . . . , N ).

Using the bootstrap samples we can calculate,

the mean the bias

1
B
b = x̄ x0
x̄ = xi Bias corrected solution!
B i=1
xbc = x0 b

the variance, 95% confidence intervals,

B
1 x[1] x[2] ··· x[N ]
2
= (xi x̄)2
B i=1
[x[0.025N +1] , x[0.975N +1] ] 95% Confidence Interval

38

THE BOOTSTRAP IS VERSATILE

The beauty of the bootstrap is that it can be applied to ANY type of data errors
(not just Gaussian), and ANY function of data (not just the mean).
Input Output

Data Function of the Result

y = (y1 , y2 , y3 , . . . , yn ) x = (x1 , x2 , . . . , xm )
y data x

The number of bootstrap samples needed depends on the number of data, n, and type
of output, x. Rule of thumb: 1000 for a quick look, more for the details.

There are many codes available for advanced bootstrap implementations in statistical
programming languages like S-Plus, SAS, SPSS and Minitab, or you can write your own.

APPLIED RESAMPLING METHODS - COURSE NOTES (2001, ANU)

MICHAEL MARTIN, STEVEN STERN (AVAILABLE AS PDF FROM SPEAKER)
http://rses.anu.edu.au/~malcolm/Teaching/MC/

39
 BOOTSTRAP PRACTICAL

The trajectory of a cannonball

1
y(t) = m1 + m2 t m3 t 2
2
99 %
m2
Data
95 %

68 %
Bootstrap
(yi , ti )
Elevation (m)

solutions

Best fit
solution

Time (s) m1

1
ri = yi m01 m02 ti + m03 t2i
2
1 0 2
yjb = rjb + m01 m02 tj m t
2 3 j

Bootstrap resampling of residuals gives new data sets and from these we get Bootstrap solutions.

40

MONTE CARLO SIMULATION:

APPLICATIONS
USED IN A ‘ DIZZINGLY DIVERSE NUMBER OF WAYS’.
EXAMPLES INCLUDE

MOLECULAR CHEMISTRY: CLASSICAL BOTLZMANN, QUANTUM

MECHANICAL ENERGIES,...

USED TO SIMULATE UNCERTAINTY IN FINANCE; RISK

MINIMIZATION IN INSURANCE

MEDICAL STATISTICS: DRUG TESTING, TREATMENT PROTOCOLS

A WORKHORSE APPROACH IN THEORETICAL PHYSICS

A WORKHORSE METHOD FOR BAYESIAN INFERENCE

41

MONTE CARLO
BASED METHODS FOR DATA
FITTING

42
 FITTING MODELS TO DATA
Suppose we have a set of unknowns x = (x1 , x2 , x3 , . . . , xm )

which we want to constrain by fitting some data y = (y1 , y2 , y3 , . . . , yn )

Given a set of values for the unknowns we have a way of making predictions of
the corresponding data yp = f (x)

Suppose we have a measure of how well the predictions fit the data, e.g.
n
(yi yp,i )2
⇥(x) = 2
i=1 i

We want to find the values of the unknowns, x which best fit the data, and so the
data fit problem becomes an optimization problem

Data fit problem Optimization problem

y -> x x

43

EASY AND DIFFICULT

OPTIMIZATION PROBLEMS
A common process in nearly all areas of the sciences is to fit models to data

We represent our model as a set of unknowns and try to find their values which
fit observations and other criteria well.
.. field of nonlinear optimization

Deterministic methods Monte Carlo methods

yp = f (x)

Data fit
Uses Derivatives
Un
kn
ow

1
wn
n2

kno Derivative Free

Un Infrasound array tuning (Kennett et al. 2003)

...and sometime not so easy

Some times the process is easy

44

MONTE CARLO METHODS FOR

GLOBAL OPTIMIZATION
Random sampling (> 1968)
Exhaustive and
random search

Simulated annealing (> 1982)

...Parallel Tempering
Genetic algorithms and
Evolutionary computation
(> 1992)

Ants, Bees and particle swarms...

Neighbourhood algorithm (> 1999)

MONTE CARLO METHODS IN GEOPHYSICAL INVERSE PROBLEMS

SAMBRIDGE, M., AND MOSEGAARD, K., REV. GEOPHYS., 40, DOI:10.1029/2000RG000089 5 DECEMBER 2002.

45
 UNIFORM SEARCH
IN NESTED GRIDS

Parameter space

Nested grid sampling for three unknowns

There will be one grid dimension for every unknown

This is simple in low dimensions but expensive in many dimensions

46

PARALLELIZED SEARCH IN THE HYPER CUBE

USING THE HILBERT CURVE
1 bit 2 bits

2b NODES PER AXIS LYNDMEYER

RECURSIVE
CONSTRUCTION
2nb NODES IN TOTAL ALGORITHM

3 bits 5 bits

ilab Software

A SPACE FILLING CURVE THAT PROVIDES THE OPTIMAL MAPPING OF A

REGULAR LATTICE ON TO A LINE.

47

PARALLELIZED SEARCH IN THE HYPER CUBE

USING THE HILBERT CURVE

A SPACE FILLING CURVE THAT PROVIDES THE OPTIMAL MAPPING OF A

REGULAR LATTICE ON TO A LINE.

48
 PARAMETER SEARCH
PRACTICAL
Searching for the minimum of the Rosenbrock function

Uniform nested
Rosenbrock function
grid sampling

f (x1 , x2 ) = (1 x1 )2 + 100(x2 x21 )2

Finding parameters that provide best fit to data misfit function is complex

49

RANDOM SEARCH
IN GEOPHYSICS

Random generation of 1-D whole Earth wavespeed profiles to fit

seismic travel times Eigenfrequencies, Earth’s Mass and MoI.

Press (1968)

Keilis-Borok and Yanovskaya (1967) Valius (1968)

50

MONTE CARLO METHODS AND

PARALLEL COMPUTATION

MC methods are ideally suited to

modern computational hardware
x2

x1

Multiple evaluations of the data fit can be

spread across different processors

51
 RANDOM SAMPLING OF THE
HYPERSPHERE
2r
n Nc Vns /Vnc
Curse of
2 1024 78%
dimensionality
4 1.05E6 31%
10 1.3E30 0.2%
16 1.5E48 3.6E-4%

Uniform random samples in hypersphere Gaussian density in n-dimensions

Sampling density as a function of radius

1
Uniform
0.8

Sampling density
0.6

0.4
Gaussian

0.2

0
0 0.5 1 1.5 2 2.5 3
Radius r
N s = 804 N g = 1024
Estimated T̃nsp 2n(b 1) n/2 T̃ngp N
= = ilab Software
compute time t̄ np ( n2 + 1) t̄ np

52

NEIGHBOURHOOD ALGORITHM

A derivative free method for multi-dimensional search and optimization .

The Neighbourhood Algorithm was designed to have as few tunable parameters

as possible and yet achieve a robust self adaptive search behaviour.
ilab Software

Motivation
Given the value of the objective function at a set of (random) positions in
parameter space, how can we make use of the ensemble as a whole to determine
where to sample next ?

One answer is to use the position of each sample as

Unknown 2

representative of its neighbourhood in parameter space

Let previous sampling guide the search

Geophysical Inversion with a Neighbourhood Algorithm -I. Searching a parameter space

Unknown 1 Sambridge, M., Geophys. J. Int., 138, 479-494, 1999.

53

HOW TO DEFINE
NEIGHBOURHOODS

From Okabe et al. (1995)

Nearest neighbour regions are Voronoi cells (Descarte, 1644, Voronoi 1908).

Also known as Dirichlet domains, Thiessen polygons, Wigner-Seitz cells.

Mathematics, Meterology and Crystalography...

54
 VORONOI CELLS ARE
EVERYWHERE

55

NA EXPLAINED

Neighbourhood algorithm explained with the aid of

three blind mice

Lets see how they run...

56

A RESTRICTED RANDOM WALK

Conceptually simple

1. Take n random walks inside m ‘better’ cells

2. Build new Voronoi cells and go to 1

57
 NA EXPLAINED
A RANDOM WALK INSIDE VORONOI CELLS

1 2 3

4 5 6

Cheese

1. Take n random walks inside m ‘better’ cells

2. Build new Voronoi cells and go to 1

58

NA EXAMPLE

A self adaptive search process

Searching a 2-D landscape

Mt. Kosciuszko

59

NA EXAMPLE: INFRASOUND

From Kennett et al. (2003)

A PROBLEM WITH 2 UNKNOWNS

60
 NA EXAMPLE: SEISMIC RECEIVER
FUNCTIONS

A PROBLEM WITH 24 UNKNOWNS

61

A COMPARISION OF METHODS

62

MONTE CARLO ENSEMBLES

Fitting 1-D seismic models to receiver functions

Searching
for best fit
solutions

Searching
for all
acceptable
fit solutions

63
 BAYESIAN
TRANS-DIMENSIONAL
INFERENCE
& RESEARCH TRENDS

64

A PROBABILISTIC FRAMEWORK
Likelihood

p(m|d) = C p(d|m) p(m)

Prior

Probabilities are associated with inferences

model estimation <-> model uncertainty

The theory of probabilities is at bottom nothing but common sense reduced to calculus.
Laplace, Théorie analytique des probabilités, 1820

65

EXAMPLE: MEASURING THE MASS OF AN OBJECT

Before After

p(m|d) = C p(d|m) p(m)

Gaussian distribution

1 1 (m m̄)
2

p(m) = e 2 2

1 1
10.0)2
Where we start Prior PDF p(m) = e 2 (m
2
1 2(m 11.2)2
Measure m = 11.2 ± 0.5 Likelihood p(d = 11.2|m) = e
0.5 2
1 1 (m 10.96)
2
Where we end Posterior PDF p(m|d = 11.2) = e 2 1/5

66
 BAYESIAN INFERENCE PRACTICAL

What size of profit can be made from the lottery ?

Given numbers of winners in each division of a lottery, calculate how many tickets were sold.

i 1 2 3 4 5 6
di 14 169 3059 149721 369543 802016
pi 1
8145060 678756 36696 732 300 144

Find an answer by

1. A simple Frequentist approach

3. A Bayesian Inference approach

5. A Maximum Likelihood approach

p(m|d) = C p(d|m) p(m)

67

SAMPLING AND OPTIMIZATION

Optimization Probabilistic sampling

framework framework

Maximum

Markov chain
Parameter Monte Carlo sampling
space
Gradient optimization

Data fit or Log Likelihood function Samples from posterior PDF

⇥(m) = ||d g(m)||22 + 2

||m||22 p(m|d) = k p(d|m) p(m)

68

MARKOV CHAIN MONTE CARLO

Target PDF p(m|d) = C p(d|m) p(m) M-H ALGORITHM

Proposal PDF q(m2 |m1 )

m1 m2

M-H ⇥
Acceptance p(m2 |d)q(m1 |m2 )
=1 EXAMPLE

probability p(m1 |d)q(m2 |m1 )

TARGET

Issues: REJECTION
Efficiency ALGORITHM

Dimension
M-H
ALGORITHM

Tempering, Hamiltonian sampling,...

69
 SIMULATION OF A MULTI-MODAL
PROBABILITY DISTRIBUTION
A sum of two Gaussians

70

SIMULATION OF A MULTI-MODAL
PROBABILITY DISTRIBUTION

71

TRANSDIMENSIONAL INVERSION

Variable dimension

p(m, k|d) = C p(d|m, k) p(m|k)p(k)

⇥
p(d|m2 , k2 )p(m2 |k2 )p(k2 )q(m1 , k1 |m2 , k2 )
=1 |J|
p(d|m1 , k1 )p(m1 |k1 )p(k1 )q(m2 , k2 |m1 , k1 )

Birth-Death McMC algorithm, RJ-McMC algorithm

Geyer & Moller (1994) Green (1995)

k2 = k1 ± 1 k
m(x) = mj j (x)
j=1

72
 EXAMPLE REGRESSION
' Make everything as simple as possible, but not simpler’
Einstein
Birth-Death example 60
1-D Regression Data
40 Partition modelling
20

Recover the discontinuous function d 0

(gray) from the noisy data (red) mT = (f , x, k)T
−20 True model position
variable
height
−40 variable

−60
0 2 4 6 8 10
Properties of the ensemble x
may be examined

TRUE
Uncertainty in
ENSEMBLE
AVERAGE
model dimension
DATA

ilab Software

73

REGRESSION
WITH TRANSDIMENSIONAL MCMC
Dynamically fit the height and the position of discontinuity

Bayesian sampling
of partition models

N=6 N=11
Mean solution

True solution

Data driven smoothing !

No explicit regularization

Self adaptive
N=8 N=3

Examine robust properties of the ensemble of

candidate solutions
74

MCMC PRACTICAL
1-D regression problems with multiple partitions and unknown noise

Original signal and data Bayesian sampling results Estimating data noise

Data (xi , yi )

75
 1-D SPATIAL INFERENCE

m(z)

m1
z1

z i-2
Depth z

i-1 th layer
mi-1
z i-1
i-th layer
mi
hi

zi

DIFFERENT LAYERED CONDUCTIVITY MODELS

ALL FIT AEM DATA TO WITHIN ERROR

76

INVERSION FOR 1-D EARTH MODELS

Using Airborne EM to constrain
subsurface conductivity mT = (c, x, k)T
Conductivity
0.01 0.1
0

10 Uncertainty in
discontinuity position
20

30
Depth (m)

40

50

60

70

80

Courtesy
R. Brodie, M. Hartley
ilab Software

77

WHAT IF WE DO NOT KNOW THE

LEVEL OF DATA NOISE ?
When data uncertainty is uncertain
N
⇥
1 1 ⇤ |d g(m)|2
L(d|m) = ⇤ exp ⇥ ⇥
(⇥ 2 )N 2 i ⇥2
⇥
1 1 Cd ( 1, 2)
L(d|m) = exp (d g(m))T Cd 1 (d g(m))
(|Cd |(2 )N )1/2 2
We can solve for
data noise correlation too

p(m, i , k|d) =C L(d|m, i , k) p(m|k)p( i )p(k)

Markov chain, (m, i , k) (m , i, k )

⇥
p(m , ⇥i , k )|d)q(m, ⇥i , k|m , ⇥i , k )
=1 |J|
p(m, ⇥i , k)|d)q(m , ⇥i , k |m, ⇥i , k)

78
 NOISE AND MODEL COMPLEXITY
BOTH VARIABLE
N
⇥
1 1 ⇤ |d g(m)|2
L(d|m) = ⇤ exp
( ⇤ 2⇥)N 2 2
i
⇤2

Transdimensional with fixed data uncertainty Transdimensional with variable data uncertainty
mT = (f , x, k)T and k mT = (f , x, , k)T
σestimated = 4 are related
A) 60
40 Model complexity and data noise variable
20
Model too complex A) 60

0 Data over fit Recovered model is

−20
d0 accurate
−40
0 2 4 6 8 10
−60
σestimated = 10
60
0 2 4
x 6 8 10

B)
40
B)
20 Model complexity OK 1
Estimated Model
Data fit good
0
complexity is accurate
−20

−40
0 2 4 6 8 10 0
0 10 20 30 40 50
σestimated = 30 Number of Partitions
60
C) C) 1
40 TRUE
20
Estimated Data ESTIMATED
Model too smooth noise is accurate
0 DATA
Data under fit
−20
0
ilab Software −40 5 10 15 20 25
0 2 4 6 8 10 σ

79

GEOCLIMATE EXAMPLE
80˚ 100˚ 120˚

mT = (fi , i , x, k)
T
East Asian Winter Monsoon

40˚ 40˚

Hongyuan
Data from 6m core of Hongyuan peat bog located on eastern Qinghai Tibetan plateau (From Large et al. 2009)
East Asian Summer
Monsoon
20˚ 20˚

Indian Summer Monsoon Geochemical proxies for constraining Holocene evolution of the East Asian
monsoon
80˚ 100˚ 120˚

DATA RESULTS
HongyuanH
2
Total C (g/int./m2)

6000
1

Use partition modelling to 4000

0

constrain the depths (timing) of

2000
−1
0

abrupt changes in climatic −2

Density (g/cm3)

conditions 0.5
−3

0 100 200 300 400 500 600

Depth (cm)
Hongyuan DelC
0 3

30 2
δ13C(°/°°)

28
Data noise differs between 26
1

different data types

0
24
Nitrogen

−1
4
N (wt %)

−2
2
−3
Probability for400change
Effect of abrupt changes not
0 100 200 300 500 600
Depth (cm)
0

always in the same direction

Hydrogen Carbon

1
60
C (wt %)

Probability

40

Better
p(change point)

20

0 Poorer
Do not know number of 6

change points
H (wt %)

4
0
2 0 100 200 300 400 500 600
Depth (cm)

0
0 100 200 300 400 500 600
ilab Software Depth (cm) From Gallagher et al. (2010)

80

JOINT INVERSION OF MULTIPLE

CLASSES OF DATA
Receiver function data
a)
Receiver function (noise)
Receiver Function

Receiver function (no noise)

Posterior PDF on the

mT = (v, i , x, k)
T

number of layers
Noise magnitude and correlation length
-5 0 5 10 15 20 25 30
Time (s)
b)
4.0
Dispersion data
3.5
Group velocity

3.0

2.5 Surface wave dispersion (noise)

Surface wave dispersion (no noise)
2.0
20 40 60 80 100
Period (s)

Marginal density map on PDF on

a) subsurface velocity b)
changepoints c) d) e) f)

Receiver function Dispersion Both

data only data only data

After Bodin et al. (2012)

81
 CLASSIC TOMOGRAPHY:
SINGLE MODELS WITH A FIXED GRID
Damping

110˚
−10˚
120˚ 130˚ 140˚ 150˚
−10˚
110˚
−10˚
120˚ 130˚ 140˚ 150˚
−10˚ Tikhonov
−20˚ −20˚ −20˚ −20˚
solutions
−30˚ −30˚ −30˚ −30˚

Smoothing
−40˚ −40˚ −40˚ −40˚

120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚

110˚ 110˚

120˚ 130˚ 140˚ 150˚

110˚

−10˚ −10˚
120˚ 130˚ 140˚ 150˚
110˚ 120˚ 130˚ 140˚ 150˚
110˚
−10˚ −10˚
−10˚ −10˚
−20˚ −20˚

−30˚ −30˚

−20˚ −20˚
−20˚ −20˚

−40˚ −40˚

120˚ 130˚ 140˚ 150˚

110˚

−30˚ −30˚
−30˚ −30˚

50 x 35
−40˚ −40˚
−40˚
B-splines nodes
−40˚

120˚ 130˚ 140˚ 150˚

110˚ 120˚ 130˚ 140˚ 150˚
110˚

120˚ 130˚ 140˚ 150˚ 130˚ 140˚ 150˚

110˚ 120˚
110˚
−10˚ −10˚ −10˚ −10˚

2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4

−20˚ −20˚ −20˚ −20˚

min ||⇤d G⇤m||2 + ||m mo ||2 + ⇥||⇤m||2

−30˚ −30˚ −30˚ −30˚
m
−40˚ −40˚ −40˚ −40˚
After Bodin et al. (2009)
120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚
110˚ 110˚

82

2-D SPATIAL INFERENCE

o
0 0
o o
0

o
9 S o
9 S o
9 S

o
18 S o
18 S o
18 S

o o
27 S 27 S o
27 S

o o
36 S 36 S o
36 S

o o
o o
120 E 130 E 140oE 150 oE o o
120 E 130 E 140oE 150 oE o o
130 E 140oE
45 S110o E 45 S110o E o
45 S110o E 120 E 150 oE

20 uniformly random 2 overlapping Voronoi meshes 4 overlapping Voronoi meshes

Voronoi cells

o
0

o
9 S
Birth-Death McMC with unknowns mT = (f , x, k)T
o
18 S
Point by point averaging of output gives an ensemble mean, covariance
o
27 S
measures.

Effective mesh resolution increases with number of meshes -> super

o
36 S

o o o
130 E 140oE
resolution.
45 S110o E 120 E 150 oE

20 Voronoi meshes

83

2-D TOMOGRAPHY
THE WISDOM OF THE CROWD

Individual models along the Markov chain

mT = (s, x, k)T
120˚ 130˚ 140˚ 150˚
110˚ 120˚ 130˚ 140˚ 150˚
120˚ 130˚ 140˚ 150˚ 110˚
110˚ −10˚ −10˚
−10˚ −10˚
−10˚ −10˚

−20˚ −20˚ −20˚

−20˚ −20˚
−20˚

−30˚ −30˚ −30˚ −30˚

−30˚ −30˚

−40˚ −40˚ −40˚ −40˚ Best data −40˚ −40˚

fit model
120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚
120˚ 130˚ 140˚ 150˚ 110˚ 110˚
110˚

20 cells 50 cells 30 cells

64

Optimization solution 62
Number of cells
1750 cells + regularization Ensemble mean
140˚
60
vs
Number of Cells

130˚ 150˚
iterations
120˚ 130˚ 140˚ 150˚
110˚ 120˚ 58
110˚
−10˚ −10˚ −10˚
−10˚
56

54

−20˚ −20˚ −20˚ 52

−20˚

50

48
0 0.5 1 1.5 2 2.5 3 3.5 4
−30˚ −30˚ −30˚ −30˚
Iterations x 10
5

−40˚ −40˚ −40˚ −40˚ Estimated PDF

for the
110˚
120˚ 130˚ 140˚ 150˚
110˚
120˚ 130˚ 140˚ 150˚
number of cells
120˚ 130˚ 140˚ 150˚ 130˚ 140˚ 150˚
110˚ 120˚
110˚
−10˚ −10˚ −10˚ −10˚

−20˚ −20˚ −20˚ −20˚ 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4
−30˚ −30˚ −30˚ −30˚

0 10 20 30 40 50 60 70 80
Number of cells

Super resolution & self regularization

−40˚ −40˚ −40˚ −40˚

120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚

110˚ 110˚

True Data

84
 Uncertainty estimates for free

Ensemble average Ensemble variance

120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚
110˚ 110˚
−10˚ −10˚ −10˚ −10˚

−20˚ −20˚ −20˚ −20˚

−30˚ −30˚ −30˚ −30˚

−40˚ −40˚ −40˚ −40˚

120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚

110˚ 110˚

0.0 0.1 0.2 0.3 0.4

2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4

120˚ 130˚ 140˚ 150˚ 130˚ 140˚ 150˚

110˚ 120˚
110˚
−10˚ −10˚ −10˚ −10˚

−20˚ −20˚ −20˚ −20˚

−30˚ −30˚ −30˚ −30˚

−40˚ −40˚ −40˚ −40˚

120˚ 130˚ 140˚ 150˚ 120˚ 130˚ 140˚ 150˚

110˚ 110˚

True Data

85

AMBIENT NOISE IMAGING

Transdimensional inversion with unknown data noise 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4

mT = (v, i , x, k)
T
1 120˚ 130˚ 140˚ 150˚
110˚
−10˚ −10˚
p(n|d)

120˚ 130˚ 140˚ 150˚

110˚ 0
−10˚ 0 200 400 600 800 1000 1200 1400 1600 1800 2000 −20˚ −20˚
−10˚
Number of Cells
1 1 1
p(λ|d)

p(a|d)

p(b|d)

−20˚ −20˚
−30˚ −30˚

0 0 0
0.2 0.3 0.4 0.6 0.8 1 1.2 0 0.5 1 1.5
−30˚ −30˚ λ a b
−40˚
PDF Data uncertainty
−40˚

−40˚ −40˚ 120˚ 130˚ 140˚ 150˚

110˚

120˚ 130˚ 140˚ 150˚

110˚
0.0 0.1 0.2 0.3 0.4 0.5 0.6
0.0 0.1 0.2 0.3 0.4 0.5 0.6

140˚ 150˚ Velocity cell density Output: Error in the ensemble mean
120˚ 130˚ 140˚ 150˚
110˚ 130˚ 140˚
120˚ 150˚
−10˚ −10˚ 110˚
−10˚ −10˚

−20˚ −20˚
−20˚ −20˚

−40˚ −40˚ −30˚ −30˚

−30˚ −30˚

−40˚ −40˚
140˚ 150˚ −40˚
−40˚

120˚ 130˚ 140˚ 150˚

110˚
120˚ 130˚ 140˚ 150˚
110˚

ilab Software
From Bodin et al. (2011)
86

AUSMOHO EXAMPLE
Constraining the Moho depth using seismic Reflections, Refraction, receiver
function, tomography, historical seismicity and gravity (oceans)

87
 AUSMOHO EXAMPLE
Constraining the Moho depth using seismic Reflections, Refraction, receiver function,
tomography, historical seismicity and gravity (oceans)

Multiple data sets with unknown relative uncertainty

88

BODY WAVE IMAGING OF THE CMB

Transdimensional tomography vs Linearized inversion

Lowermost 300 km of the mantle

Linear inversion
with fixed parametrization
Tkalčić et al. (2002)

Transdimensional
Probabilistic sampling
(mobile in position
and variable in complexity)

Long wavelength features are similar

Short scale structure resolved Data: PKPab-df and PcP-P differential travel times From Young et al. (2012)

89

SOME PUBLICATIONS ON TRANS-DIMENSIONAL INVERSION

Electrical resistivity inversion
Malinverno, A., 2002. Parsimonious Bayesian Markov chain Monte Carlo inversion in a
107 nonlinear geophysical problem, Geophys. J. Int., 151(3), 675–688.

Malinverno, A. & Briggs, V., 2004. Expanded uncertainty quantification in inverse problems: Hierarchical Bayes and empirical Bayes,
Geophysics, 69, 1005.

Climate histories
Hopcroft, P., Gallagher, K., & Pain, C., 2007. Inference of past climate from borehole temperature data using Bayesian Reversible Jump
Markov chain Monte Carlo, Geophys. J. Int., 171(3), 1430–1439.

Hopcroft, P., Gallagher, K., & Pain, C., 2009. A Bayesian partition modelling approach to resolve spatial variability in climate records from
borehole temperature inversion, Geophys. J. Int., 178(2), 651–666.

Thermochronology
Gallagher, K., Charvin, K., Nielsen, S., Sambridge, M., & Stephenson, J., 2009. Markov chain Monte Carlo (MCMC) sampling methods to
determine optimal models, model resolution and model choice for Earth Science problems, Marine and Petroleum Geology, 26(4), 525–535.

Stratigraphic modelling
Charvin, K., Gallagher, K., Hampson, G., & Labourdette, R., 2009. A Bayesian approach to inverse modelling of stratigraphy,
part 1: method, Basin Research, 21(1), 5–25.

Geochronology
Jasra, A., Stephens, D., Gallagher, K., & Holmes, C., 2006. Bayesian mixture modelling in geochronology via Markov chain Monte Carlo,
Mathematical Geology, 38(3), 269–300.

Geochemistry
Gallagher, K., Bodin, T., Sambridge, M., Weiss, D., Kylander, M., & Large, D., 2011. Inference of abrupt changes in noisy geochemical records
using Bayesian Transdimensional changepoint models, Earth and Planetary Science Letters, doi:10.1016/j.epsl.2011.09.015.

90
 SOME PUBLICATIONS ON TRANS-DIMENSIONAL INVERSION

Seismology, Regression, boreholes, wirelines and AEM

Sambridge, M., Gallagher, K, Jackson, A and Rickwood, P., 2006. Trans-dimensional inverse problems, Model Comparision and the Evidence,
Geophy. J. Int., 167, 528-542, doi: 10.1111/j.1365-246X2006.03155.x, 2006.

Bodin, T. & Sambridge, M., 2009. Seismic tomography with the reversible jump algorithm,
Geophys. J. Int., 178(3), 1411–1436.

Bodin, T., Sambridge, M., and Gallagher, K., 2009. A self-parameterising partition model approach to tomographic inverse problems,
Inverse Problems, 25, 055009, doi:10.1088/0266-5611/25/5/055009.

Sambridge, M., Bodin, T., Reading, A., and Gallagher, K., 2010. Data inference in the 21st Century: Some ideas from outside the box ,
Australian Soc. of Exploration Geophysics 21st International Geophysics Conference and Exhibition, Extended Abstracts, 22-26 August, Sydney,
Australia.

Reading, A., Bodin, T., Sambridge, M., Howe, S. and Roach, S., 2010. Down the borehole but outside the box: innovative approaches to
wireline log data interpretation , Australian Soc. of Exploration Geophysics 21st International Geophysics Conference and Exhibition,
Extended Abstracts, 22-26 August, Sydney, Australia.

Reading, A. M., Cracknell, M. J., Sambridge, M., and Forster, J. G., 2011. Turning geophysical data into geological information, or, why a broader
range of mathematical strategies is needed to better enable discovery, Preview,, 140, 24-29, doi:10.1071/PVv2011b151p24.

T. Bodin, M. Sambridge, H. Tkalcic, P. Arroucau, K. Gallagher,and N. Rawlinson, 2012. Transdimensional Inversion of Receiver Functions
and Surface Wave Dispersion, J. Geophys. Res., 117, B02301. doi:10.1029/2011JB008560.

Brodie, R. C. and Sambridge, M., 2012. Transdimensional Monte Carlo Inversion of AEM Data
Australian Soc. of Exploration Geophysics 22nd International Geophysics Conference and Exhibition,
Extended Abstracts, 26-29 February, Brisbane, Australia.

Bodin, T., Sambridge, M., Rawlinson, N., and Arroucau, P., 2012. Transdimensional tomography with unknown data noise,
Geophys. J. Int., 189, 1536-1556, doi: 10.1111/j.1365-246X.2012.05414.x.

91

SUMMARY OF TRANSDIMENSIONAL INFERENCE

Ensemble based approaches which use the data to constrain both

parametrization and data noise look promising in a range of applications
in the geosciences. A general methodology.

Ensemble properties are reasonably robust and can be used as an

estimator of the solution. Uncertainty for free.
Make use of the wisdom of the crowd.

Need efficiency in higher dimensions, more flexible basis functions,

Spherical geometries.

92

MONTE CARLO

We have seen that Monte Carlo methods are at the essence computational
techniques that use (pseudo) random numbers

Applications to

Evaluation non random integrals

Simulation of random processes,

e.g. uncertainty propagation

Fitting parameters to data (optimization)

Probabilistic sampling for inference

93
 MONTE CARLO PRACTICALS

Four practicals that require the student to implement an algorithm using an

MC based approach
(1,1)

R1

1. Monte Carlo integration example

y

x = -0.6 (0,0) x x = 0.6

R2

99 %
m2 95 %
R3

(-1,-1)
68 %
Bootstrap
solutions

2. Bootstrap example
Best fit
solution

m1

3. Parameter search

4. Probabilistic inference toy problem

5. Probabilistic inference regression

6. Probabilistic inference regression with discontinuities in the data

94